[4Yin2] Interactive session 2

[4Yin2-29] Super-Resolution of Asteroid Images using GAN

〇Rintarou Kishi¹, Ren Hashimoto¹, Kimikazu Tanase¹, Rie Honda², ONC Team Hayabusa2 (1.Kochi Unversity, 2.Ehime University)

Keywords:super resolution, GAN, asteroid image, exploration

Landing surveys are becoming an important tool in the exploration of planets and small bodies. For this purpose, it is necessary to extract regions with few obstacles from low-resolution images, and super-resolution is useful. The images of planets and small celestial bodies contain various terrains and rocks with a wide size distribution. In addition, the scale of the acquired images varies widely, and the apparent features vary greatly depending on the sunlight conditions. In this study, we aim to construct a general-purpose super-resolution system for images with such characteristics using the adversarial generation network GAN. Here, SRGAN is actually applied to images of Ryugu acquired by the asteroid explorer Hayabusa2 under various resolutions and lighting conditions. We examined the accuracy of the recovery by using SRGAN trained on artificially degraded images and original images. In a preliminary experiment, the PSNR of the generated image was 29.1 dB for a 30 cm/pixel image under the same lighting conditions, which was equivalent to the analytical method. On the other hand, it was confirmed that the performance tended to be the same for low-resolution images with mixed illumination conditions, as long as the training data was also mixed, suggesting that GAN-based super-resolution is useful for images of planets and small bodies.